On the use of deep learning for ocean SAR image semantic segmentation
With water covering 71% of the surface of the Earth, and most meteorological processes stemming from the oceans, their observation is primordial to enhance our comprehension of the Earth system, improve meteorological models, and prevent hazards. Since ERS-1 (launched in 1991), C-Band Synthetic Aper...
Main Author: | |
---|---|
Other Authors: | , , , , , , , , , |
Format: | Doctoral or Postdoctoral Thesis |
Language: | English |
Published: |
HAL CCSD
2022
|
Subjects: | |
Online Access: | https://theses.hal.science/tel-04030623 https://theses.hal.science/tel-04030623/document https://theses.hal.science/tel-04030623/file/2022IMTA0327_Colin-Aurelien.pdf |
Summary: | With water covering 71% of the surface of the Earth, and most meteorological processes stemming from the oceans, their observation is primordial to enhance our comprehension of the Earth system, improve meteorological models, and prevent hazards. Since ERS-1 (launched in 1991), C-Band Synthetic Aperture Radar (SAR) has beenused to observe the ocean surfaces. This particular electromagnetic band is especially useful for deriving information on waves, winds, precipitation, sea ice, and more at meso- and sub-mesoscale. The subject of this thesis is the segmentation, or pixel-per-pixel classification, of the ocean surface C-Band SAR observations. The generation of segmentation maps is possible through the use of machine learning frameworks that are able to extract information from the large data volume produced by the satellites Sentinel-1A and Sentinel-1B, which were launched in 2014 and 2016 as part of ESA’s Copernicus program. These observations are collocated with third-party sensors (ground stations, buoys, satellite-boarded instruments), manually annotated segmentations, or meteorological models to be able to train deep learning models and ensure their capacity through extensive tests. These studies show promising uses of new SAR-derived information and propose guidelines for building dedicated segmentation datasets and models. Nombreux sont les phénomènes météorologiques prenant naissance dans les océans, dont 71% de la surface de la Terre est couverte. L’observation des étendues marines est primordiale pour accroître notre compréhension du système Terre, améliorer les modèles météorologiques et atténuer l’effet des catastrophes naturelles. Depuis le lancement d’ERS-1 en1991, les radars à synthèse d’ouverture (SAR, d’après l’acronyme anglais) en bande C sont utilisés pour observer les surfaces océaniques. La bande C est, en effet, particulièrement utile pour obtenir des informations sur les vagues, le vent, les précipitations, la banquise, que ce soit à méso- ou à sous-méso échelle. La thèse ci-présente ... |
---|